23. Oktober 2018 / Marc Toussaint

Best Paper Award for Prof. Marc Toussaint

What we hear from newspapers, industry or politicians suggests that Deep Learning is the answer to AI.

But Intelligence is so much more than training models on big data. AI was, and perhaps should continue to be, a research agenda to understand, formally describe and synthesize the whole variety of fascinating behaviors that we subsume under the term \emph{intelligence} in humans and animals.

Geometric and physical reasoning is one such aspect. Animals are surprisingly skilled in using tools and solving physical problems. The big inventions and innovations in humankind were conceived not by mimicking big data, but by generalizing from previous experience and literally imagining new things. In Cognitive Science, the term \emph{intuitive physics} subsumes the idea that people permanently utilize some basic physics understanding when interpreting and
perceiving our world, but also when anticipating and making decisions. This is shown already in small children. We all are animals evolved to interact with our physical world. And seemingly we are equipped with a powerful engine in our mind for exploiting the possibility of such interaction, for geometrically imagining things, anticipating physical effects, and reasoning about what can be done.

Until recently there were hardly satisfying ideas of what kind of computational engine could perform such physical reasoning. Classical AI planning is super strong on abstract problems, but it was hard to including geometric and physical aspects in such reasoning. Planning methods that emerged from robotics research are strong in reasoning about the motion of robots themselves, but the possibilities of physical with all objects was hard to include. And physics simulations are very powerful to forward-simulate the physics of large scenes. But a simulation only computes one future---it does not envision a plethora of possible futures, it does not compute or estimate all the things that might be possible and desirable if only one would choose appropriate actions.

Prof. Toussaint from University of Stuttgart and his collaborators from MIT recently won the best paper award at the international flagship robotics conference RSS---for an algorithm that does exactly that. Hundreds and thousands of geometric and physical constellations are considered, evaluated, and perhaps investigated further on whether they
make up a consistent and desirable future. Thereby, this "physical AI engine" scans systematically, but in some sense also creatively, through that abundance of geometric and physically possible futures and pick those consistent to goals.

A key component of the system is utilizing constrained optimization algorithms. These are algorithms that continuously modify solutions until they're consistent. And they are extremely powerful to stretch and squeeze potential geometric and physical futures until they in fact fulfill the laws of physics. The keyword behind this is \emph{differentiable physics}, which is a key technology for optimization algorithms to modify solutions until they're physically correct.

The second key component is how to systematically scan through the possibilities that come from your own actions. Here, the algorithm utilizes classical logic---but this is a logic about what optimization problems should even be considered. So, the logic here is for making clever decisions about what is fed to the optimization algorithms below. Together they make a physical reasoning process that aims to create a whole variety of possible futures that are geometrically and physically consistent and lead to a given goal.

The demonstrations given in the paper might seem childish for a human: A ball is thrown and hit with a stick to end up at a goal. A stick is used to kick a ball to enable reaching for it. But it is fascinating that an algorithm can come up with such solutions only on based knowing about geometric shapes and physics. Now the race is on to make the algorithm faster, exploit all the computation power computers have, and maybe end up with a physical AI engine that will outperform human physical thinking and problem solving. Whether they could be equally creative in conceiving of novel physical machines and constructions is another question.